Pith. sign in

REVIEW 4 major objections 4 minor 45 references

SteerPose: Simultaneous Extrinsic Camera Calibration and Matching from Articulation

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Mentally rotating 2D poses, as a learned network does for a hypothesized camera rotation, jointly yields extrinsic camera calibration and cross-view instance matches, with no calibration tools or precomputed correspondences.

desk verdict Novel and plausible pipeline for joint extrinsic calibration and cross-view matching from articulation, but the real-data comparisons are overclaimed and the retry-based success gate plus orthographic training assumption need scrutiny before the method is fully trusted. read the letter →

arxiv 2506.01691 v2 pith:QCKDIACU submitted 2025-06-02 cs.CV

Sang-Eun Lee , Ko Nishino , Shohei Nobuhara This is my paper
classification cs.CV
keywords extrinsiccameracalibrationcross-viewinstancematching2Dposerotationrotation-covariantnetworkSinkhorngeometricconsistencylossmarkerlessmotioncaptureclass-agnosticmodel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SteerPose tries to establish that the moving subjects themselves—a person walking, pigs in a pen, a cheetah on the savanna—are enough to calibrate a multi-camera rig and to decide which instance in one view is which in another. The paper's bet is that articulation is a general, class-agnostic prior: a network that learns to 'mentally rotate' a 2D pose into another view, conditioned on a hypothesized relative rotation, transfers across human and animal skeletons. At test time the network stays frozen, and only the rotation is optimized, with differentiable Sinkhorn matching supplying cross-view instance correspondences and a geometric consistency loss requiring that the current rotation admit a valid relative translation. If the claim holds, markerless multi-view 3D pose capture becomes practical in uncontrolled settings, with no checkerboards, wands, or manual correspondence, and novel animal species can be reconstructed from off-the-shelf 2D pose estimates alone.

What carries the argument

SteerPose is a small Transformer that encodes each 2D joint into 32-dimensional tokens with hierarchical positional encodings and appends a rotation token built from the Rodrigues vector of $R$; self-attention lets it model both local joint relationships and global pose structure, which helps under partial occlusions. The rotation-covariant property $f(g(\cdot)) = g'(f(\cdot))$ is what carries the geometric content: it says the network's output behaves like the projection of a pose rotated in 3D. Two losses drive the inference-time optimization. The matching loss uses Sinkhorn's soft assignment, with dummy targets for occlusion asymmetry and bidirectional evaluation for cycle consistency, to score how well $Q(R)$ explains $P'$. The geometric consistency loss $L_{geom}$ builds the linear system $A[x_1 \ldots x_N \; t]^{\top} = 0$ from collinearity and coplanarity constraints and penalizes the singular-value ratio $\sigma_1/\sigma_2$, using closed-form derivatives, so that the optimized rotation admits a valid relative translation; back-propagation then refines $R$ while SteerPose stays frozen.

What would settle it

Render synthetic 3D poses with perspective cameras at shrinking camera-to-subject distances while holding the relative rotation fixed, and track SteerPose's 2D prediction error and the recovered rotation error; if both degrade sharply as perspective distortion grows, the orthographic training data, rather than the geometric optimization, is carrying the result. A complementary check is to run the class-agnostic model on a species far outside the quadruped training distribution, such as a bipedal bird—a category for which the paper reports no class-agnostic result—and see whether calibration and matching success fall to the level of the five-point baseline.

Watch

Extended reading notes

Core claim

The paper's central claim is that extrinsic camera rotation and cross-view instance correspondences can be recovered simultaneously, and purely from 2D poses of freely moving articulated subjects, by a rotation-covariant network combined with differentiable matching. SteerPose is trained in a fully supervised way on synthetic view pairs rendered from 3D pose data, so that for a 3D-to-2D projection $f$ and a 3D rotation $g$ it realizes the companion transform $g'$ satisfying $f(g(\cdot)) = g'(f(\cdot))$: given a 2D pose and a relative rotation, it predicts the pose as seen from the other view. At inference the network is frozen and the rotation $R$ is the only optimized variable; Sinkhorn matching soft-assigns the transformed poses $Q(R)$ to the observed poses $P'$, a bidirectional matching loss scores the assignment, and the new geometric consistency loss stacks collinearity and coplanarity equations into a linear system so that the estimated $R$ guarantees a valid solution for the relative translation. The paper shows that this joint optimization escapes false matches that satisfy 2D pose similarity alone, and that pairwise calibration integrated by motion averaging and bundle adjustment delivers accurate multi-camera poses and 3D reconstructions across humans, quadrupeds, and birds, with a class-agnostic model trained on 40 quadruped species reconstructing novel animals such as cheetahs and pigs.

Load-bearing premise

The claim rests on a transfer assumption: a rotation-covariant mapping trained on orthographically projected synthetic views of 3D poses will still produce reliable rotated 2D poses for real perspective cameras, noisy off-the-shelf keypoint detections, and species with different skeleton topologies; if that transfer fails, the matching and the geometric loss are fed bad inputs and the joint optimization cannot find the true rotation.

Editorial extensions

If this is right

  • Multi-camera motion capture becomes deployable without a calibration stage: the subjects' own articulation supplies the calibration target, so rigs can be set up in livestock pens, wildlife sites, or homes and calibrated from the footage itself.
  • Cross-view instance identity and camera rotation are solved by one optimization, and the geometric loss explicitly prevents the two from being trapped in a 2D-similarity false match that has no valid translation solution.
  • A single class-agnostic SteerPose trained on 40 quadruped species transfers to unseen animals, with a keypoint-completion step that fills missing or topologically absent joints by applying identity rotations; the same completion also fills keypoints lost to occlusion.
  • The method beats the five-point RANSAC baseline on every tested species and is comparable to a state-of-the-art image-matching pipeline while being more reliable in wide-baseline, texture-poor scenes where image features fail.
  • Triangulated 3D poses from the calibrated views can serve as pseudo ground truth for training monocular 3D pose estimators of novel animals, so a raw multi-view recording of an unstudied species becomes a training dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same rotate-and-match loop should work when the camera moves and the subject stays (egocentric or handheld video): the data needed is just 2D pose sequences plus a hypothesized relative rotation, so sequential calibration from a single moving camera is a natural extension the paper does not explore.
  • Because training uses orthographic projection, the method's accuracy should degrade with perspective distortion; a testable prediction is that prediction error of $Q(R)$ and the final rotation error grow as camera-to-subject distance shrinks, marking where close-range setups would need perspective-aware training.
  • Chaining the rotation-covariant mapping across frames could couple instance tracking with calibration in one differentiable loop, something the paper leaves to the Sinkhorn module per view pair rather than over time.
  • The pose-space matching makes SteerPose usable as a differentiable prior inside larger optimizations, for example jointly refining all cameras and correspondences of a multi-camera cluster at once instead of the pairwise-plus-motion-averaging pipeline the paper adopts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes SteerPose, a transformer that takes a 2D pose and a relative camera rotation R as input and predicts the 2D pose as seen from the other view, trained purely on synthetically rendered orthographic projections of 3D pose data. At inference, SteerPose is kept frozen while R is optimized by backpropagation through a differentiable Sinkhorn matching loss and a novel geometric consistency loss, so that extrinsic rotation and cross-view instance correspondences are estimated jointly. The method is evaluated on in-the-wild datasets of humans and several animal species, compared against a five-point RANSAC baseline and learning-based methods (LightGlue, MASt3R), and extended to multi-view calibration via motion averaging and bundle adjustment. A class-agnostic variant is claimed to generalize to unseen quadruped species and to enable 3D pose reconstruction of novel animals.

Significance. The central idea is attractive and original: instead of relying on static scene features or calibration tools, the method treats articulated bodies as structured calibration targets and learns a mental-rotation operator on 2D poses. If validated, this would make multi-camera calibration and correspondence search practical in uncontrolled, texture-poor, wide-baseline scenarios, and the proposed geometric consistency loss is a sensible way to prune false correspondences that satisfy 2D similarity but are geometrically inconsistent. The paper also demonstrates a potentially useful application of generating pseudo-ground-truth 3D poses for novel species. However, the current experimental validation has three load-bearing gaps: the retry protocol in D.3 gates reported results, the orthographic-to-perspective transfer is unanalyzed, and the class-agnostic evaluation in Appendix A uses SteerPose itself for keypoint completion. Because these issues affect the central claim of real-world validity, the paper needs substantial revision before it can be accepted.

major comments (4)
  1. [D.3] The evaluation protocol is success-gated. The text states that calibration is retried with a new random initial rotation up to five times whenever the reprojection error exceeds 10 pixels. Because this success criterion is essentially the same quantity as the reported E2D metric and is also used in the multi-view pipeline, the reported results in Tables 1, 2, and 6 are effectively best-of-five outcomes with an unstated failure rate. Please report the number of attempts actually used per sequence, the fraction of sequences that fail all five attempts, and results without any retry (or with a fixed number of optimization steps regardless of the 10-pixel threshold). Without this, the comparison against MASt3R and LightGlue, which do not have such a gate, is not apples-to-apples.
  2. [D.1, Section 5.1] The orthographic-to-perspective transfer is a load-bearing assumption that is never tested. SteerPose is trained only on orthographic projections of 3D poses, yet at inference it consumes 2D keypoints from perspective cameras. Under perspective projection, the same 2D pose can arise from many different 3D configurations, so the transformed pose Q(R) is not a function of the 2D pose and R alone; the learned mapping can transfer only if perspective effects are negligible or implicitly absorbed. The paper does not specify how 2D keypoints are normalized (centering, scale, or focal-length handling) and does not ablate the effect of perspective distortion. Please add a controlled synthetic experiment where the same 3D poses are rendered with perspective cameras of varying focal length and subject depth, and report how calibration accuracy degrades with these parameters. This is essential to support the claim that the method works on real perspective images.
  3. [Appendix A, Table 4] The class-agnostic generalization results are partly circular. To evaluate SteerPose on Beagle Dog, whose skeleton topology differs from Animal3D, the authors use SteerPose itself with identity rotation to fill in missing joints before computing the test loss. This means the reported class-agnostic test errors (e.g., 0.16 for Beagle Dog) measure the model's self-consistency under identity rotation rather than true generalization to an unseen skeleton topology. The calibration results in Table 1 that use this completed keypoint input are thereby confounded. Please report class-agnostic results without model-based keypoint completion, or use a fixed, independent skeleton mapping (e.g., a manual joint correspondence or a separately trained completion module), and state explicitly which rows of Tables 1 and 6 rely on completion.
  4. [5.2, Table 2] The claim that SteerPose achieves accuracy 'comparable to' MASt3R is not supported by the table on the real-image rows. For Beagle Dog, Toddler, and Volleyball, MASt3R reports AUC@20 values of 0.94, 0.89/0.90, and 0.97, while SteerPose reports 0.47, 0.66, and 0.72, respectively. These are large gaps, not small differences. The accompanying text also states that SteerPose is 'more robust in challenging scenarios,' but Table 2 shows the opposite on the most realistic rows. Please rewrite this comparison to acknowledge the performance gap, discuss the failure modes on real off-the-shelf keypoints, and, ideally, separate the influence of the retry protocol from genuine algorithmic robustness.
minor comments (4)
  1. [3.1, Eq. (1)] The dissimilarity D(q_i(R), p'_i) is used before being defined; please state explicitly that it is the L2 distance between the 2D joint coordinates and clarify whether the poses are centered or normalized before computing this distance.
  2. [3.2] The sentence about dummy targets for N ≠ N' cites Sinkhorn and Knopp [34], which is the algorithm reference, but the dummy-augmentation procedure is not described; please provide a concrete description and an appropriate citation.
  3. [Appendix A, Table 4] The table uses asterisks for class-agnostic without completion, but the meaning of the asterisk is only explained in the text and not in the table caption; please add a footnote to the table and also state which class-agnostic variant was used in Tables 1 and 6.
  4. [D.2] The architecture description states that the rotation is embedded as a 32-dimensional token and also that the input includes 'three parameters representing the rotation by the Rodrigues vector'; please clarify how the 3-vector is converted into the 32-dimensional token and how it is combined with the pose tokens.

Circularity Check

1 steps flagged · score 3.0 of 10

One self-referential generalization evaluation in Appendix A; central calibration benchmarks against external ground truth are not circular.

  1. self definitional [Appendix A, Table 4 (Generalizability of SteerPose)]
    "To evaluate the class-agnostic SteerPose trained with Animal3D using the test set from Beagle Dog, we used SteerPose itself to align the skeleton structure by applying identity rotation to the 2D poses. That is, by masking the missing 2D joints in the input, rotating 2D poses to the same viewpoint by SteerPose can estimate the 2D keypoint positions of the masked joints. This can also fill in missing keypoints due to occlusions. This “keypoint completion” step can enhance the generalizability of the class-agnostic model as shown in the two “Class-agnostic” columns in Table 4."

    The class-agnostic generalization loss in Table 4 is computed on Beagle Dog test poses after missing joints (e.g., the Animal3D neck) have been estimated by SteerPose itself under identity rotation. Because the completed keypoint positions are outputs of the same network being evaluated, those joints have no independent ground truth; the reported average keypoint errors on the completed skeleton partly measure the model’s agreement with its own hallucinations. This is a self-definitional evaluation: the target used to score generalization is defined using the model’s own prediction. The result is presented as evidence that SteerPose generalizes effectively to unseen species and supports the class-agnostic claims, so it is load-bearing.

full rationale

The central calibration-and-matching claim is not circular: SteerPose is pretrained on synthetically projected 3D poses independent of the test cameras, the inference-time rotation R is optimized against external views, and the headline results in Tables 1, 2, 3, and 6 are scored against ground-truth camera poses and instance correspondences from established datasets. The comparisons with the five-point method, LightGlue, and MASt3R are external benchmarks. The one genuine self-referential element is Appendix A: to bridge the skeleton mismatch between Animal3D and Beagle Dog, the authors use SteerPose itself to complete missing joints before measuring generalization error, so the evaluation target for those joints is partly generated by the model under test. This weakens the novel-animal generalization evidence but does not invalidate the main calibration results, which rely on external ground truth. The retry gate reported in Appendix D.3 (up to five attempts with a 10-pixel reprojection success criterion) is a selection-bias concern rather than a circular reduction, so it is not counted as a circular step here.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central method is a learned network trained on synthetic orthographic projections, an inference-time optimization, and an evaluation protocol. The main premises are transferability of the learned 2D-pose rotation to real perspective cameras and unseen species, the hand-chosen loss weighting, and the retry-based success gate.

free parameters (4)
  • lambda, geometric loss weight = not reported
    Sets the balance in L = Lmatch + lambda * Lgeom (Sec. 3.2). The paper says it is empirically determined but does not give the value.
  • alpha, Sinkhorn similarity scale = 3
    Defines similarity s = 2 / (1 + exp(alpha * Lkp)) in Sec. 3.2. Chosen by hand and controls how confident the soft assignment is.
  • Retry threshold and maximum attempts = 10 px, 5 attempts
    Appendix D.3 declares calibration successful only below 10 px reprojection error and retries up to 5 times. This hand-chosen gate affects which runs enter the reported metrics.
  • Training data synthesis settings = 100 cameras, 20 rolls, 3000 pairs
    Appendix D.1 fixes these sampling counts. They affect the diversity of rotations SteerPose sees and therefore its accuracy, but they are not fitted.
assumptions (6)
  • domain assumption Orthographic projection is an adequate proxy for perspective cameras when training SteerPose.
    Training pairs are generated with orthographic projection (Appendix D.1) while test cameras are perspective (Sec. 5.1).
  • domain assumption Body articulation is transferable across quadruped species and across small skeleton topology differences.
    The class-agnostic model trained on Animal3D is applied to Cheetah, Bama Pig, and Beagle Dog; topology mismatches are patched by keypoint completion (Appendix A).
  • domain assumption Off-the-shelf 2D pose estimators deliver keypoints accurate enough for matching and geometric constraints.
    Real experiments depend on SuperAnimal, RTMO, and HRNet detections (Table 7); no sensitivity analysis to detector noise is provided.
  • standard math Minimizing sigma1/sigma2 is a valid objective for finding a rotation that admits a translation solution.
    The collinearity/coplanarity linear system must be rank deficient for a solution; this is a necessary condition from epipolar geometry [35], but the paper's wording 'guarantees' is stronger than the loss provides.
  • ad hoc to paper Gradient descent on R from random initializations finds a usable optimum.
    The inference-time optimization has no convergence guarantee; the paper mitigates with retries (Appendix D.3).
  • standard math The Sinkhorn soft assignment is a reliable differentiable proxy for discrete matching.
    The matching module in Sec. 3.2 uses Sinkhorn [9] to compute soft correspondences; its convergence to a doubly stochastic matrix is standard, but its suitability for pose matching is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SteerPose: Simultaneous Extrinsic Camera Calibration and Matching from Articulation." pith.science (2026). https://pith.science/paper/QCKDIACU

@misc{pith2026250601691,
  author       = {Pith},
  title        = {Pith review of: SteerPose: Simultaneous Extrinsic Camera Calibration and Matching from Articulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QCKDIACU}},
  note         = {Machine review of arXiv:2506.01691}
}
read the original abstract

Can freely moving humans or animals themselves serve as calibration targets for multi-camera systems while simultaneously estimating their correspondences across views? We humans can solve this problem by mentally rotating the observed 2D poses and aligning them with those in the target views. Inspired by this cognitive ability, we propose SteerPose, a neural network that performs this rotation of 2D poses into another view. By integrating differentiable matching, SteerPose simultaneously performs extrinsic camera calibration and correspondence search within a single unified framework. We also introduce a novel geometric consistency loss that explicitly ensures that the estimated rotation and correspondences result in a valid translation estimation. Experimental results on diverse in-the-wild datasets of humans and animals validate the effectiveness and robustness of the proposed method. Furthermore, we demonstrate that our method can reconstruct the 3D poses of novel animals in multi-camera setups by leveraging off-the-shelf 2D pose estimators and our class-agnostic model.

Figures

Figures reproduced from arXiv: 2506.01691 by the authors.

Figure 1
Figure 1. Camera calibration by SteerPose. Given a set of 2D poses P of targets with known articulation, SteerPose predicts their appearance under a relative camera rotation R as Q(R) which can be matched against 2D poses P ′ observed in another view in a differentiable manner. Novel matching and geometric losses evaluate the validity of R as well as the matching between Q(R) and P ′ , and their gradients are back-propagated … view at source ↗
Figure 2
Figure 2. SteerPose learns to rotate an input 2D pose pi to the 2D pose qi(R) observed from another view at relative rotation R in a fully-supervised manner. SteerPose serves as a function g ′ (·) that performs mental rotation to form a rotation-covariant transformation of 3D poses under projection f(·) satisfying f(g(·)) = g ′ (f(·)) where g(·) is the rotation in 3D by R. The training dataset was synthesized by rendering 3D … view at source ↗
Figure 1
Figure 1. Suppose we have two sets of 2D poses P = {pi} (i = 1, . . . , N) and P ′ = {p ′ i ′} (i ′ = 1, . . . , N′ ) of the target estimated from images captured at different viewpoints C and C ′ . By specifying a relative rotation R from C, the pretrained SteerPose transforms each 2D pose pi ∈ P of C to those at another view at rotation R as qi(R) ∈ Q(R), and Sinkhorn algorithm can find the best correspondences between the … view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: Two-view extrinsic calibration and matching results. The ground-truth and estimated camera poses are shown [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 4
Figure 4. Figure 4: Visualization for multi-view calibration results. The ground-truth cameras are in black, the initial camera [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Our method refines camera poses through the nonlinear optimization of reprojection errors within a unified [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 5
Figure 5. Figure 5: Two-view extrinsic calibration and matching results. Colored lines across the views represent the estimated [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Visualization for multi-view calibration results. The inputs (left) show detected 2D poses of multiple [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Network architecture of SteerPose Train Calibration (Synthetic) Calibration (Real) Cheetah [39] Jules flick1 (20190309) Romeo flick (20190227) Romeo flick (20190227) Bama Pig [3] 0–1400 1400–1750 0–1750 (HRNet) Beagle Dog [3] 0–80 80–112 0–112 (SuperAnimal [41]) Pigeon…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 37 canonical work pages

  1. [1]

    3d-pop-an automated annotation approach to facilitate markerless 2d-3d tracking of freely moving birds with marker-based motion capture

    Hemal Naik, Alex Hoi Hang Chan, Junran Yang, Mathilde Delacoux, Iain D Couzin, Fumihiro Kano, and Máté Nagy. 3d-pop-an automated annotation approach to facilitate markerless 2d-3d tracking of freely moving birds with marker-based motion capture. In CVPR, pages 21274–21284, 2023

    work page 2023

  2. [2]

    3d-muppet: 3d multi-pigeon pose estimation and tracking

    Urs Waldmann, Alex Hoi Hang Chan, Hemal Naik, Máté Nagy, Iain D Couzin, Oliver Deussen, Bastian Goldluecke, and Fumihiro Kano. 3d-muppet: 3d multi-pigeon pose estimation and tracking. IJCV, pages 1–18, 2024

    work page 2024

  3. [3]

    Three-dimensional surface motion capture of multiple freely moving pigs using mammal

    Liang An, Jilong Ren, Tao Yu, Tang Hai, Yichang Jia, and Yebin Liu. Three-dimensional surface motion capture of multiple freely moving pigs using mammal. Nature Communications, 14(1):7727, 2023

    work page 2023

  4. [4]

    Balan, and Michael J

    Leonid Sigal, Alexandru O. Balan, and Michael J. Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. IJCV, 87:4–27, 2010

    work page 2010

  5. [5]

    Using deeplabcut for 3d markerless pose estimation across species and behaviors

    Tanmay Nath, Alexander Mathis, An Chi Chen, Amir Patel, Matthias Bethge, and Mackenzie Weygandt Mathis. Using deeplabcut for 3d markerless pose estimation across species and behaviors. Nature protocols, 14(7): 2152–2176, 2019

    work page 2019

  6. [6]

    Multi-animal pose estimation, identification and tracking with deeplabcut

    Jessy Lauer, Mu Zhou, Shaokai Ye, William Menegas, Steffen Schneider, Tanmay Nath, Mohammed Mostafizur Rahman, Valentina Di Santo, Daniel Soberanes, Guoping Feng, et al. Multi-animal pose estimation, identification and tracking with deeplabcut. Nature Methods, 19(4):496–504, 2022

    work page 2022

  7. [7]

    Mental rotation of three-dimensional objects

    Roger N Shepard and Jacqueline Metzler. Mental rotation of three-dimensional objects. Science, 171(3972): 701–703, 1971

    work page 1971

  8. [8]

    Rotation equivariant vector field networks

    Diego Marcos, Michele V olpi, Nikos Komodakis, and Devis Tuia. Rotation equivariant vector field networks. In ICCV, Oct 2017

    work page 2017

Show all 45 references
  1. [9]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. NeurIPS, 26, 2013

    2013

  2. [10]

    A flexible new technique for camera calibration

    Zhengyou Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330–1334, 2000

    2000

  3. [11]

    Garrido-Jurado, R

    S. Garrido-Jurado, R. Muñoz-Salinas, F.J. Madrid-Cuevas, and M.J. Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition, 47(6):2280–2292, 2014

    2014

  4. [12]

    Wand-based multiple camera studio calibration

    Joel Mitchelson and Adrian Hilton. Wand-based multiple camera studio calibration. CVSSP Technical Report, 2003

    2003

  5. [13]

    Panoptic studio: A massively multiview system for social interaction capture

    Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE TPAMI, 2017

    2017

  6. [14]

    Marker- less motion capture with unsynchronized moving cameras

    Nils Hasler, Bodo Rosenhahn, Thorsten Thormahlen, Michael Wand, Jürgen Gall, and Hans-Peter Seidel. Marker- less motion capture with unsynchronized moving cameras. In CVPR, 2009

    2009

  7. [15]

    Global structure-from-motion revisited

    Linfei Pan, Dániel Baráth, Marc Pollefeys, and Johannes L Schönberger. Global structure-from-motion revisited. In ECCV, 2024

    2024

  8. [16]

    Structure-from-motion revisited

    Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016

    2016

  9. [17]

    David G. Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 60(2):91–110, November 2004

    2004

  10. [18]

    Visual slam algorithms: A survey from 2010 to 2016

    Takafumi Taketomi, Hideaki Uchiyama, and Sei Ikeda. Visual slam algorithms: A survey from 2010 to 2016. IPSJ Transactions on Computer Vision and Applications, 9(1):16, 2017

    2010

  11. [19]

    Extrinsic camera calibration from a moving person

    Sang-Eun Lee, Keisuke Shibata, Soma Nonaka, Shohei Nobuhara, and Ko Nishino. Extrinsic camera calibration from a moving person. IEEE RA-L, 7(4):10344–10351, 2022. doi:10.1109/LRA.2022.3192629

    2022

  12. [20]

    Human pose as calibration pattern; 3d human pose estimation with multiple unsynchronized and uncalibrated cameras

    Kosuke Takahashi, Dan Mikami, Mariko Isogawa, and Hideaki Kimata. Human pose as calibration pattern; 3d human pose estimation with multiple unsynchronized and uncalibrated cameras. In Proc. CVPRW, 2018. 8 SteerPose: Simultaneous Extrinsic Camera Calibration and Matching from A...

    2018

  13. [21]

    Wide-baseline multi-camera calibration using person re-identification

    Yan Xu, Yu-Jhe Li, Xinshuo Weng, and Kris Kitani. Wide-baseline multi-camera calibration using person re-identification. In CVPR, 2021

    2021

  14. [22]

    Camera calibration using parallel line segments

    Gaku Nakano. Camera calibration using parallel line segments. In Proc. ICPR, pages 1505–1512, 2021. doi:10.1109/ICPR48806.2021.9412183

    2021

  15. [23]

    Nicola Garau, Francesco G. B. De Natale, and Nicola Conci. Fast automatic camera network calibration through human mesh recovery. Journal of Real-Time Image Processing, 2020

    2020

  16. [24]

    Spatiotemporal multi-camera calibration using freely moving people

    Sang-Eun Lee, Ko Nishino, and Shohei Nobuhara. Spatiotemporal multi-camera calibration using freely moving people. IEEE Robotics and Automation Letters, 10(5):4818–4825, 2025

    2025

  17. [25]

    Superpoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPRW, pages 224–236, 2018

    2018

  18. [26]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, pages 4938–4947, 2020

    2020

  19. [27]

    Lightglue: Local feature matching at light speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, pages 17627–17638, 2023

    2023

  20. [28]

    Steerers: A framework for rotation equivariant keypoint descriptors

    Georg Bökman, Johan Edstedt, Michael Felsberg, and Fredrik Kahl. Steerers: A framework for rotation equivariant keypoint descriptors. In CVPR, pages 4885–4895, 2024

    2024

  21. [29]

    End2end multi-view feature matching with differentiable pose optimization

    Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In ICCV, pages 477–487, 2023

    2023

  22. [30]

    Fischler and R

    M. Fischler and R. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. CACM, 24(6):381–395, 1981

    1981

  23. [31]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024

    2024

  24. [32]

    Grounding image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r. arXiv:2406.09756, 2024

    2024 arXiv

  25. [33]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017

    2017

  26. [34]

    Concerning nonnegative matrices and doubly stochastic matrices

    Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343–348, 1967

    1967

  27. [35]

    R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000

    2000

  28. [36]

    Matrix backpropagation for deep networks with structured layers

    Catalin Ionescu, Orestis Vantzos, and Cristian Sminchisescu. Matrix backpropagation for deep networks with structured layers. In ICCV, 2015

    2015

  29. [37]

    Combining two-view constraints for motion estimation

    Venu Madhav Govindu. Combining two-view constraints for motion estimation. In CVPR, 2001

    2001

  30. [38]

    Fast and robust multi-person 3d pose estimation and tracking from multiple views

    Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. IEEE TPAMI, 44(10):6981–6992, 2022

    2022

  31. [39]

    Acinoset: A 3d pose estimation dataset and baseline models for cheetahs in the wild

    Daniel Joska, Liam Clark, Naoya Muramatsu, Ricardo Jericevich, Fred Nicolls, Alexander Mathis, Mackenzie W Mathis, and Amir Patel. Acinoset: A 3d pose estimation dataset and baseline models for cheetahs in the wild. In ICRA, pages 13901–13908, 2021

    2021

  32. [40]

    Egohumans: An egocentric 3d multi-human benchmark

    Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh V o, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. arXiv preprint arXiv:2305.16487, 2023

    2023 arXiv

  33. [41]

    Superanimal pretrained pose estimation models for behavioral analysis

    Shaokai Ye, Anastasiia Filippova, Jessy Lauer, Steffen Schneider, Maxime Vidal, Tian Qiu, Alexander Mathis, and Mackenzie Weygandt Mathis. Superanimal pretrained pose estimation models for behavioral analysis. Nature communications, 15(1):5165, 2024

    2024

  34. [42]

    RTMO: Towards high-performance one-stage real-time multi-person pose estimation, 2023

    Peng Lu, Tao Jiang, Yining Li, Xiangtai Li, Kai Chen, and Wenming Yang. RTMO: Towards high-performance one-stage real-time multi-person pose estimation, 2023

    2023

  35. [43]

    An efficient solution to the five-point relative pose problem

    David Nistér. An efficient solution to the five-point relative pose problem. IEEE TPAMI, 26(6):756–770, 2004

    2004

  36. [44]

    Animal3d: A comprehensive dataset of 3d animal pose and shape

    Jiacong Xu, Yi Zhang, Jiawei Peng, Wufei Ma, Artur Jesslen, Pengliang Ji, Qixin Hu, Jiehua Zhang, Qihao Liu, Jiahao Wang, et al. Animal3d: A comprehensive dataset of 3d animal pose and shape. In ICCV, pages 9099–9109, 2023

    2023

  37. [45]

    class-specific

    Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, pages 106–115, 2024. 9 SteerPose: Simultaneous Extrinsic Camera Calibration and Matching from Articulation A PREPRINT Appendices Appendix A General...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.